Contagion Networks: Evaluator Bias Propagation in Multi-Agent LLM Systems
Researchers have developed a framework called Contagion Networks to quantify how biases spread within multi-agent systems where large language models act as evaluators. Experiments using DeepSeek-chat demonstrated that evaluator biases consistently propagate between agents, even when using the same underlying model. The study identified mitigation strategies, such as increasing the size of the evaluator committee, which significantly reduced bias propagation. AI
IMPACT Provides a method to understand and mitigate bias propagation in multi-agent LLM systems, crucial for reliable AI deployment.